Official Discrepancy Corpus

The official repository preserves a dedicated diagnostic report explaining predicate by predicate why the current Algorithm 1 rejects four curves: 62a1, 66b1, 105a1, and 141c1. These four curves all passed semistability, isogeny exclusion, ramification, optimality, the rank-zero gate, the 2-descent gate for $E'$, and the current $\operatorname{BSD}(E,2)$ gate, but each failed on four specific predicates: $\operatorname{ord}_2 L^{\mathrm{alg}}=-2$ (CLZ20 requires $-1$), the 2-torsion structure of $E'(\mathbb Q)_{\mathrm{tors}}$ is $(2,2)$ (failing the required cyclic condition), $f'(x_0)$ is the square of a rational number, and the $\mathcal S$ non-emptiness gate failed. The document explicitly positions this data as an adversarial regression corpus for the theorem-router, and provides a precise, directly quotable discipline: if a future version suddenly accepts these four curves, the first tag applied should be REGRESSION?, not NEW BSD BREAKTHROUGH! — this is exactly applying the discipline of "reason=OPEN, do not guess" from document 08 in reverse to the scenario of "unexpectedly passing."